Physical Human-Robot-Human Interaction

• Physical Human-Robot-Human Interaction (pHRHI) is an emerging subfield that extends contact-based interactions to triadic scenarios where a robot mediates physical collaboration among multiple humans.
• Key methodologies include latent-objective inference, adaptive admittance control, and multi-modal sensing techniques such as capacitive servoing and tactile skin integration.
• Experimental paradigms span collaborative manipulation, rehabilitation, and manufacturing, highlighting challenges in contact attribution, role arbitration, and real-time multimodal feedback.

Searching arXiv for relevant pHRHI and closely related pHRI foundations. Physical Human-Robot-Human Interaction (pHRHI) denotes interaction settings in which a robot is physically coupled to more than one human, or physically mediates interaction between humans, so that force exchange, contact, compliance, intent inference, and safety must be handled at the level of a multi-agent physical system rather than a dyadic human-robot pair. In the current literature, pHRHI is less a fully standardized subfield than an emerging synthesis of contact-based pHRI, collaborative manipulation, assistive robotics, rehabilitation, and human-human haptic negotiation. A recurring theme across this body of work is that physical contact should not be treated solely as disturbance or collision; it is also a channel for latent task communication, coordination, and role allocation, provided that sensing, planning, and control can resolve ambiguity, preserve safety, and remain legible to all participants (Losey et al., 2021, Rysbek et al., 2023, Farajtabar et al., 2024).

1. Conceptual scope and relation to contact-based pHRI

The most direct conceptual framing of pHRHI in the supplied literature is triadic: a robot may mediate physical collaboration among multiple humans, or two humans may physically interact with or through a robot while the robot must infer shared or conflicting intentions (Losey et al., 2021). This differs from standard contact-based pHRI, which is usually formulated around one human physically interacting with one robot. The broader survey literature on contact-based pHRI classifies physical interaction into direct contact, indirect interaction through an intermediary object, and proximity interaction; all three are relevant to pHRHI, but triadic systems add multi-human intent, multi-contact attribution, and mediated force transmission as central issues rather than edge cases (Farajtabar et al., 2024).

A consistent implication across the corpus is that most available technical foundations for pHRHI come from neighboring literatures rather than from explicit pHRHI formulations. The latent-objective inference model of intentional correction during pHRI assumes one human physically correcting one robot during an ongoing task (Losey et al., 2021). Safe close-contact architectures based on tactile skin and peripersonal space are likewise described for dyadic human-robot interaction (Švarný et al., 2018). Multi-modal intent inference with Gaussian-process force models, capacitive limb servoing, adaptive admittance for contact-rich manufacturing, and language-grounded assistive trajectory editing all remain primarily dyadic (Haninger et al., 2021, Erickson et al., 2021, Niaz et al., 2024, Wang et al., 15 Jan 2026). Even so, these works repeatedly identify components that transfer directly to pHRHI: latent-intent estimation, selective learning from physical correction, pre-contact and post-contact safety layers, mode-aware compliant control, and explicit transparency about what the robot has inferred.

The pHHI literature strengthens this interpretation by showing that collaborative object manipulation often cannot be understood from motion alone. In dyadic tray transport with uncertain or conflicting destination goals, force exchange functions as negotiation over intent, commitment, and initiative (Rysbek et al., 2023). This suggests that pHRHI should not be reduced to “multi-contact safety”; it is also a problem of physical communication among multiple agents under shared constraints.

2. Physical interaction as communication and latent intent inference

One of the clearest technical shifts relevant to pHRHI is the treatment of physical correction as an observation about a hidden objective rather than merely as a disturbance. In the latent-objective formulation of pHRI, robot state $x$, robot action $u_r$, and human action $u_h$ evolve under

$\dot{x} = f(x, u_r + u_h),$

while the human is assumed to optimize a reward

$$r(x,u_r,u_h ; \theta) = \theta \cdot \phi(x,u_r,u_h) - \lambda \|u_h \|^2,$$

with latent parameter $\theta$ encoding the task objective (Losey et al., 2021). The problem is posed as a POMDP in which $u_h$ is an observation of the hidden objective, and belief over $\theta$ is updated online from physical corrections. The practical result is that after contact ends, the robot does not simply return to the old plan; it replans the remainder of the task using an updated estimate of the intended objective.

This framework is directly relevant to pHRHI because it formalizes a principle that appears repeatedly elsewhere in the literature: force exchange is information-bearing. In pHHI object relocation, projected force, projected velocity, and projected power reveal symbolic directional intent toward candidate goals (Rysbek et al., 2023). For each agent,

$P_k(t)=F_k \cdot v_k,$

and for a candidate goal direction $g_i$,

$u_r$0

These features are specifically proposed because raw force alone is ambiguous under locomotion, grasping, and oscillatory motion; force-kinematic coupling is more predictive of who is proposing, resisting, or conceding (Rysbek et al., 2023).

The multi-modal Gaussian-process MPC framework for collaborative assembly offers a complementary latent-mode perspective. There, the robot models human interaction wrench as a stochastic function of robot pose conditioned on a discrete task mode, performs Bayesian mode inference, and optimizes assistance over the belief distribution (Haninger et al., 2021). Although only one human is modeled, the formulation is already belief-aware and multi-modal, which is structurally close to what pHRHI would require if multiple humans induce discrete shared or conflicting modes.

A plausible implication is that mature pHRHI systems will need to combine these strands: continuous latent objectives inferred from correction (Losey et al., 2021), discrete or phase-structured modes inferred from force distributions (Haninger et al., 2021), and symbolic negotiation cues extracted from force-kinematic projections (Rysbek et al., 2023). The existing papers do not yet provide a unified triadic inference model, but they define the main representational axes along which such a model would likely be built.

3. Perception, contact sensing, and embodied communication channels

Perception in pHRHI is not limited to global human tracking. The literature distinguishes pre-contact sensing, post-contact sensing, and embodied communication interfaces, each with different technical roles. A foundational safety architecture couples artificial tactile skin with peripersonal space. In that view, tactile skin supports post-contact safety by detecting and localizing contact, while peripersonal space provides a protective safety zone for pre-contact behavior modulation, aligned with the two collaborative safety modes of power and force limitation and speed and separation monitoring (Švarný et al., 2018). The paper is conceptual rather than algorithmic, but its layered architecture is highly relevant to pHRHI, where multiple humans may approach or touch different robot body parts simultaneously.

Local body-relative sensing is exemplified by multidimensional capacitive servoing. A 6-electrode capacitive array mounted on the end effector estimates a 4D local limb pose

$u_r$1

from temporal capacitance measurements using a feedforward neural regressor, enabling a robot to move proximally or distally along limbs while adapting to pose (Erickson et al., 2021). The reported interaction primitive is dyadic and local, but it is precisely the sort of sensing layer needed when a pHRHI robot must remain aligned to one person’s limb under occlusion while another person is nearby.

Distributed intrinsic force sensing supplies a different communication substrate. A class-one spherical 6-bar tensegrity with 12 force-sensing nodes uses internal load redistribution to classify interaction modes such as null, drop, squeeze, and handle from time windows of sensor data (Barkan et al., 2021). The best reported configuration—Random Forest on abstract features with a 10-sample window—achieves accuracy $u_r$2 and AUC $u_r$3 (Barkan et al., 2021). The paper does not attempt multi-human source separation, but it demonstrates that morphology plus internal sensing can encode expressive interaction patterns without external skin.

Another distinct direction is symbolic tactile communication through intrinsic robot sensing. In pHRI-DIGI-TACT, a human writes handwritten digits on an uninstrumented touchpad attached to a robot flange, and a Bi-LSTM classifies the resulting end-effector force and moment signals online (Sinico et al., 31 Mar 2025). The multi-user dataset substantially improves generalization to unseen users, including under variable robot poses, and augmentation explicitly handles reversed and rotated inputs (Sinico et al., 31 Mar 2025). This is not triadic interaction, but it operationalizes the robot as a physical communication interface rather than only a motion platform.

Speech-conditioned contact adaptation extends the communicative repertoire further. BRIDGE allows real-time modification of assistive robot trajectories in position, velocity, and force using natural language, then provides verbal assurance or clarification in response (Wang et al., 15 Jan 2026). Trajectories are represented as

$u_r$4

and landmark-relative modifications use Gaussian weighting or artificial potential fields (Wang et al., 15 Jan 2026). A major pHRHI implication is that physical interaction often requires explicit shared understanding among multiple parties, not only compliant mechanics. The paper’s central finding—that bidirectional verbal feedback improved ratings of interactivity, grounding, and transparency—supports the broader point that mediated physical coordination requires communicative accountability, especially when subtle force or speed changes may not be reliably perceived (Wang et al., 15 Jan 2026).

4. Control and mediation architectures

The control literature relevant to pHRHI is dominated by compliant interaction, but it increasingly combines compliance with estimation, hierarchy, and anticipation. A recurring baseline is impedance or admittance control. In the latent-objective correction framework, impedance control provides safe compliant yielding during contact, while the robot simultaneously records force as evidence for learning, then replans the rest of the task after release (Losey et al., 2021). In the trajectory-deformation approach, impedance control is augmented so that physical interaction modifies not only the actual trajectory but also a finite future segment of the desired trajectory (Losey et al., 2017). The closed-form deformation rule

$u_r$5

makes physical contact a channel for future path editing rather than momentary deflection (Losey et al., 2017).

For contact-rich tasks with stiff environments, adaptive admittance has been explicitly made phase-dependent. In drilling-like pHRI, a two-layer intention recognition mechanism detects coarse subtasks—Idle, Tool-Attachment, Driving, Contact—and estimates continuous progress during Driving to trigger damping changes before contact (Niaz et al., 2024). The admittance law is

$u_r$6

with fixed mass $u_r$7 and damping switched among $u_r$8, $u_r$9, and $u_h$0 (Niaz et al., 2024). This anticipatory scheduling reduced human effort during Driving by $u_h$1 and oscillation amplitude at Contact by $u_h$2 (Niaz et al., 2024). For pHRHI, the transferable principle is that mediation should be phase-aware and predictive, not based on a single fixed compliance regime.

Hierarchical constraints complicate this further. Direction-constrained control for hierarchical pHRI introduces an angular cone around the desired task velocity,

$u_h$3

to avoid the extremes of unconstrained HQP, which can redirect motion arbitrarily, and strict task scaling, which can deadlock motion at boundaries (Xu et al., 2024). The method solves a minimum-error and a minimum-angle subproblem in parallel, then blends the results at the boundary of the directional constraint. The paired variable-admittance law adapts damping according to mismatch between desired and realized velocity near active constraints (Xu et al., 2024). This is highly pertinent to pHRHI because multi-human interaction under shared constraints will frequently require “acceptable directional compromise” rather than exact reproduction of any single user’s force direction.

Robustness to impact disturbances motivates another nonlinear compliance design. The shear-thickening fluid controller (SFC) is proposed to remain easy to drag under traction forces while resisting impact-like disturbances, inspired by shear-thickening fluid behavior (Chen et al., 3 Feb 2025). The detailed equations in the supplied text are truncated, but the stated objective is explicit: compliance under intended collaboration and resistance under unexpected impact. This maps naturally to pHRHI, where a mediator robot must not relay every disturbance transparently between people.

The survey literature ties these control directions together under broader principles of safety, compliance, and human intention orientation, emphasizing impedance, admittance, variable impedance/admittance, passivity, whole-body control, and learning-based parameter adaptation as the main contact-based pHRI control families (Farajtabar et al., 2024). A plausible implication is that pHRHI will inherit these same controller classes, but with force attribution, role-sensitive arbitration, and team-level constraints added on top.

5. Experimental paradigms and application domains

The empirical literature relevant to pHRHI spans collaborative manipulation, rehabilitation, assistive care, manufacturing, and local body-relative interaction. Collaborative relocation with uncertain goals provides one of the most direct experimental analogues. In dyadic tray transport, 16 subjects arranged into 24 dyads negotiated among three candidate goal locations with combinations of NoGoal, Soft goal, and Hard goal assignments (Rysbek et al., 2023). Negotiation times differed significantly across conditions, and projected power features clustered by intended goal direction (Rysbek et al., 2023). This is not robot-mediated, but it reveals the force-level phenomena a pHRHI robot would need to interpret if inserted into similar tasks.

Assistive and rehabilitation scenarios offer another important triadic bridge. The BRIDGE study used 18 older adults across scratching, feeding, and bathing tasks with a Stretch 3 mobile manipulator, allowing real-time speech-based modification of position, velocity, and force during physical assistance (Wang et al., 15 Jan 2026). The digit-recognition interface was demonstrated in a fruit-delivery task using a KUKA LBR iiwa 14 R820, showing that tactile symbolic input can be embedded in a robot task loop with verbal confirmation and touch-based stop/resume logic (Sinico et al., 31 Mar 2025). Capacitive servoing explicitly targets limb-relative assistive tasks such as dressing and bathing, with successful contour following around bent elbows, tilted forearms, and knees across 12 participants (Erickson et al., 2021). These lines of work do not yet instantiate pHRHI explicitly, but they define concrete assistive contexts where a caregiver, therapist, or supervisor could naturally enter the interaction loop.

Manufacturing pHRI provides phase-structured contact tasks with strong relevance to pHRHI. The adaptive admittance study used virtual spring compression and real drilling with a KUKA LBR iiwa 7, training on 800 virtual trials and testing online under unseen conditions (Niaz et al., 2024). The direction-constrained hierarchical controller was validated on a 7-DoF xArm7 in tasks involving virtual walls, velocity constraints, and a co-assembly scenario with visual servo constraints (Xu et al., 2024). These experiments show how physical guidance interacts with geometric constraints, a situation that becomes even more acute when more than one human is present.

Generative simulation extends the application range by automating large-scale synthetic data creation for contact-rich assistive pHRI. The text2sim2real pipeline generates diverse scenes, humans, and robot trajectories from high-level prompts, collecting 4000 demonstrations per task for scratching and bathing, then achieving real-world zero-shot success rates of $u_h$4 and $u_h$5 without arm motion, and $u_h$6 and $u_h$7 with arm motion (Wang et al., 9 Apr 2026). Although this remains one-human pHRI, it establishes a plausible route for future pHRHI data generation, especially where real triadic interaction data would be difficult or unsafe to collect at scale.

A consistent pattern across these domains is that directly triadic evaluation remains rare, while domain-relevant subproblems are already well developed: negotiation cues in pHHI (Rysbek et al., 2023), objective inference from correction (Losey et al., 2021), robust compliant control under contact transitions (Niaz et al., 2024), local body-relative sensing (Erickson et al., 2021), communicative transparency (Wang et al., 15 Jan 2026), and synthetic data generation for contact-rich assistive tasks (Wang et al., 9 Apr 2026).

6. Limitations, controversies, and open directions

The most important limitation of the current literature is structural: few of the cited works explicitly model more than one human, even when their authors discuss pHRHI implications. This is stated outright for the latent-objective correction framework, the tactile-skin and peripersonal-space safety architecture, the force-sensing tensegrity system, the adaptive admittance system, BRIDGE, the Gaussian-process MPC method, the sensorimotor RL framework, the digit-recognition interface, the latent periodic-dynamics model, the generative simulation pipeline, the capacitive servoing method, and the direction-constrained controller (Losey et al., 2021, Švarný et al., 2018, Barkan et al., 2021, Niaz et al., 2024, Wang et al., 15 Jan 2026, Haninger et al., 2021, Ghadirzadeh et al., 2016, Sinico et al., 31 Mar 2025, Kobayashi et al., 2021, Wang et al., 9 Apr 2026, Erickson et al., 2021, Xu et al., 2024). As a result, pHRHI remains, in large part, an extrapolative research program grounded in dyadic pHRI and pHHI evidence.

A second unresolved issue is contact attribution. Tactile skin can localize where contact occurred but not necessarily who caused it (Švarný et al., 2018). Intrinsic force sensing in a tensegrity body captures distributed interaction patterns but does not separate simultaneous contributions from two humans (Barkan et al., 2021). Standard admittance and impedance formulations generally treat external force as a single resultant. This suggests that robust pHRHI will require either additional sensing channels or latent decomposition models for multi-agent wrench attribution.

A third controversy concerns how aggressively robots should learn from physical interaction. The objective-inference literature explicitly warns that not every observed feature change is intentional. One-at-a-Time learning reduces unintended learning by updating only the feature most correlated with a correction, yielding fewer wrong-direction updates than All-at-Once learning under noisy human behavior (Losey et al., 2021). This lesson is especially acute in pHRHI, where multiple humans may produce ambiguous or mixed-feature corrections. The same issue appears in contact-rich control more broadly: reactive compliance can be safe but may ignore long-horizon communicative content, while strong adaptation can misinterpret incidental motion.

Transparency and legibility remain open problems as well. In BRIDGE, predictability did not significantly improve even when collaboration and understanding did (Wang et al., 15 Jan 2026). In the latent-objective correction study, learning improved collaboration and satisfaction, but inferential shifts were not always more legible (Losey et al., 2021). This matters in pHRHI because one person may command a change while another person feels its effect. A plausible implication is that pHRHI will require explicit multi-party feedback mechanisms, not just silent controller adaptation.

Finally, data scarcity and evaluation methodology remain limiting. The contact-based pHRI survey emphasizes that the field is still early-stage and lacks comprehensive integration across sensing, planning, control, safety, and ethics (Farajtabar et al., 2024). For pHRHI specifically, there is not yet a standard taxonomy of joint intents, a common benchmark suite for multi-human physical coordination, or an agreed set of metrics for fairness, mediated transparency, and role-sensitive safety. The generative-simulation literature suggests one possible route to large-scale training and testing (Wang et al., 9 Apr 2026), but multi-human, multi-contact scene generation is still a prospective extension rather than a demonstrated capability.

Taken together, the literature suggests that pHRHI is best understood as an emergent synthesis problem. Its technical core is already visible: physical contact as communication (Losey et al., 2021), force-kinematic negotiation features (Rysbek et al., 2023), multimodal sensing and safety layers (Švarný et al., 2018), compliant yet constraint-aware control (Niaz et al., 2024, Xu et al., 2024), communicative transparency (Wang et al., 15 Jan 2026), and scalable learning pipelines (Wang et al., 9 Apr 2026). What remains largely unsolved is their integration into explicitly multi-human physical interaction systems that can infer joint or conflicting intents, attribute contact sources, arbitrate roles, and remain safe, legible, and effective for all humans involved.